#### DMCA

## DAUBNEY, XIE: ENTROPY DRIVEN HIERARCHICAL SEARCH 1 Entropy Driven Hierarchical Search for 3D Human Pose Estimation

### Cached

### Download Links

### Citations

3647 | Histograms of oriented gradients for human detection
- DALAL, TRIGGS
- 2005
(Show Context)
Citation Context ...l likelihoods p(z j∣x j) are calculated. A part is represented by a rectangular patch with two image cues exploited, edges and color. Edge cues are extracted using a set of M overlapping HOG features =-=[5]-=- placed along the edges of the part. Each feature is represented as a single normalized histogram of the local image gradients at that location and they are combined such that p(z j∣x j)edge = 1 M ∏M ... |

809 | Pictorial structures for object recognition
- Felzenszwalb, Huttenlocher
- 2005
(Show Context)
Citation Context ... will scale to extract arbitrary 3D poses. Contrary to this, in the domain of 2D pose estimation current state-of -the-art methods have been shown capable of detecting poses that are much more varied =-=[3, 8, 14]-=-. This has been achieved using generative models built around the Pictorial Structures representation [10]. Recent work in 2D pose estimation has shown that by first clustering 2D poses detection rate... |

414 |
The representation and matching of pictorial structures
- Fischler, Elschlager
- 1973
(Show Context)
Citation Context ...f -the-art methods have been shown capable of detecting poses that are much more varied [3, 8, 14]. This has been achieved using generative models built around the Pictorial Structures representation =-=[10]-=-. Recent work in 2D pose estimation has shown that by first clustering 2D poses detection rates can be significantly improved [14, 15]. Often these clusters have a semantic meaning, for example differ... |

399 |
Pattern Recognition and Machine Learning (Information Science and Statistics
- Bishop
- 2006
(Show Context)
Citation Context ...he set of Ni samples are assigned an equal weight. The sample xn j , from which the sample xni is drawn conditioned on, is called its ancestor and this method of sampling is called ancestral sampling =-=[4]-=-. Its purpose in this work is to grow the search space out from the central part of the model, exploring regions of the pose space that have a higher likelihood. The key difference between this approa... |

257 | Humaneva: Synchronized video and motion capture dataset and baseline algorithm for evaluation of articulated human motion
- Sigal, Balan, et al.
(Show Context)
Citation Context ...lts are provided for the reprojected 2D error and the 3D pose reconstruction error using different numbers of samples in Fig. 5 using the average Euclidian distance between the markers as proposed in =-=[22]-=- for 3D pose and [16] for 2D errors, where the left/right limbs are switched and the smallest reprojection error used. Note that each sample represents a hypothesized state of a single part not of the... |

256 | Recovering 3d human pose from monocular images
- Agarwal, Triggs
(Show Context)
Citation Context ...rom monocular images or video is an extremely difficult problem. Currently there are two main approaches to solving this problem, the first is to learn a direct mapping from image features to 3D pose =-=[1, 19]-=-, the second is to first extract 2D pose as an intermediate stage and then ‘lift’ this to a 3D pose [2]. The limitation with both of these approaches is that they are only applicable to poses that are... |

223 | Progressive search space reduction for human pose estimation - Ferrari, Marín-Jiménez, et al. - 2008 |

215 | Partitioned sampling, articulated objects, and interface-quality hand tracking
- MacCormick, Isard
- 2000
(Show Context)
Citation Context ...earch space is too large to be uniformly discretized and stochastic methods such as Non-Parametric Belief Propagation [21], Markov-Chain Monte Carlo [17], Variational MAP [13] or Partitioned Sampling =-=[18]-=- can be employed. A problem with theseDAUBNEY, XIE: ENTROPY DRIVEN HIERARCHICAL SEARCH 3 approaches is that they often rely on weakly constrained parts where joint locations are not forced to coincid... |

208 | Pictorial structures revisited: people detection and articulated pose estimation
- Andriluka, Roth, et al.
- 2009
(Show Context)
Citation Context ... will scale to extract arbitrary 3D poses. Contrary to this, in the domain of 2D pose estimation current state-of -the-art methods have been shown capable of detecting poses that are much more varied =-=[3, 8, 14]-=-. This has been achieved using generative models built around the Pictorial Structures representation [10]. Recent work in 2D pose estimation has shown that by first clustering 2D poses detection rate... |

188 | Tracking loose-limbed people
- Sigal, Bhatia, et al.
(Show Context)
Citation Context ...en a hypothesized root node state. For evaluation we assume that the root node of the solution is known. Contrary to many stochastic approaches, where the posterior is represented as a set of samples =-=[17, 21]-=-, ours is always represented parametrically. Samples are drawn from it to permit the update and evaluation of the posterior. A parametric model has several useful benefits. Firstly, it allows the post... |

163 |
Articulated body motion capture by stochastic search
- Deutscher, Reid
- 2005
(Show Context)
Citation Context ...tochastic approaches are iterative, this allows a fixed number of particles to efficiently search the solution space by focusing the particles into progressively smaller regions with high likelihoods =-=[7, 13, 21]-=-. A problem with these techniques is how to diffuse the particles between iterations. Often a single covariance is used which is shrunk by a predetermined fixed amount across iterations [7, 13]. In th... |

143 | Finding and tracking people from the bottom up
- Ramanan, Forsyth
- 2003
(Show Context)
Citation Context ...D pose the search space can be represented by pixel locations in the image and pose can be estimated using methods such as Dynamic Programming [8], Belief Propagation [11] or Loopy Belief Propagation =-=[20]-=-. For 3D pose estimation the search space is too large to be uniformly discretized and stochastic methods such as Non-Parametric Belief Propagation [21], Markov-Chain Monte Carlo [17], Variational MAP... |

99 | Monocular 3D pose estimation and tracking by detection
- Andriluka, Roth, et al.
- 2010
(Show Context)
Citation Context ... solving this problem, the first is to learn a direct mapping from image features to 3D pose [1, 19], the second is to first extract 2D pose as an intermediate stage and then ‘lift’ this to a 3D pose =-=[2]-=-. The limitation with both of these approaches is that they are only applicable to poses that are similar to those represented in the original training set, e.g. walking. It is unlikely they will scal... |

86 | P.: Approximating the Kullback Leibler divergence between Gaussian mixture models
- Hershey, Olsen
(Show Context)
Citation Context ...tion h(θi j) = ∫ p(xi∣θi j)ln p(xi∣θi j)dxi. Whilst there is not an exact analytical expression to calculate the Entropy of a GMM, an approximation can be used to express the GMM as a single Gaussian =-=[12]-=- ˆµi = ˆΛii = K ∑ k=1 K ∑ k=1 The Entropy can then be calculated as h(θi j) = ln λ k i jµ k i , (6) λ k ( i j Λ k ii + (µ k i − ˆµi)(µ k ) i − ˆµi) T . (7) √ (2πe) d ∣ ˆΛii∣, where d is the dimension ... |

66 | Beyond trees: common-factor models for 2D human pose recovery
- Lan, Huttenlocher
- 2005
(Show Context)
Citation Context ...the reprojected 2D error and the 3D pose reconstruction error using different numbers of samples in Fig. 5 using the average Euclidian distance between the markers as proposed in [22] for 3D pose and =-=[16]-=- for 2D errors, where the left/right limbs are switched and the smallest reprojection error used. Note that each sample represents a hypothesized state of a single part not of the entire body. For com... |

63 | Clustered pose and nonlinear appearance models for human pose estimation
- Johnson, Everingham
- 2010
(Show Context)
Citation Context ... will scale to extract arbitrary 3D poses. Contrary to this, in the domain of 2D pose estimation current state-of -the-art methods have been shown capable of detecting poses that are much more varied =-=[3, 8, 14]-=-. This has been achieved using generative models built around the Pictorial Structures representation [10]. Recent work in 2D pose estimation has shown that by first clustering 2D poses detection rate... |

36 |
A unified spatio-temporal articulated model for tracking
- Lan, Huttenlocher
- 2004
(Show Context)
Citation Context ...enerative models built around the Pictorial Structures representation [10]. Recent work in 2D pose estimation has shown that by first clustering 2D poses detection rates can be significantly improved =-=[14, 15]-=-. Often these clusters have a semantic meaning, for example different orientations of the person being detected [14]. We suggest that rather than simply clustering 2D poses a much more direct approach... |

34 | A model-based approach for estimating human 3D poses in static images
- Lee, Cohen
- 2006
(Show Context)
Citation Context ...en a hypothesized root node state. For evaluation we assume that the root node of the solution is known. Contrary to many stochastic approaches, where the posterior is represented as a set of samples =-=[17, 21]-=-, ours is always represented parametrically. Samples are drawn from it to permit the update and evaluation of the posterior. A parametric model has several useful benefits. Firstly, it allows the post... |

26 | Relevant feature selection for human pose estimation and localization in cluttered images
- Okada, Soatto
- 2008
(Show Context)
Citation Context ...rom monocular images or video is an extremely difficult problem. Currently there are two main approaches to solving this problem, the first is to learn a direct mapping from image features to 3D pose =-=[1, 19]-=-, the second is to first extract 2D pose as an intermediate stage and then ‘lift’ this to a 3D pose [2]. The limitation with both of these approaches is that they are only applicable to poses that are... |

7 | Multiple frame motion inference using belief propagation
- Gao, Shi
- 2004
(Show Context)
Citation Context ...ented by the edges. To estimate 2D pose the search space can be represented by pixel locations in the image and pose can be estimated using methods such as Dynamic Programming [8], Belief Propagation =-=[11]-=- or Loopy Belief Propagation [20]. For 3D pose estimation the search space is too large to be uniformly discretized and stochastic methods such as Non-Parametric Belief Propagation [21], Markov-Chain ... |

7 | Variational maximum a posteriori by annealed mean field analysis
- Hua, Wu
- 2005
(Show Context)
Citation Context ... For 3D pose estimation the search space is too large to be uniformly discretized and stochastic methods such as Non-Parametric Belief Propagation [21], Markov-Chain Monte Carlo [17], Variational MAP =-=[13]-=- or Partitioned Sampling [18] can be employed. A problem with theseDAUBNEY, XIE: ENTROPY DRIVEN HIERARCHICAL SEARCH 3 approaches is that they often rely on weakly constrained parts where joint locati... |

2 | Estimating 3d pose via stochastic search and expectation maximization, in
- Daubney, Xie
(Show Context)
Citation Context ...proaches is that they often rely on weakly constrained parts where joint locations are not forced to coincide [13, 21], this was recently shown to be too unconstrained for accurate 3D pose estimation =-=[6]-=-. In this work we examine how a graphical model can be utilized given a part’s state is known a priori, effectively fixing a node in the graph. To the best of our knowledge this problem has not previo... |